Who Decides What's Interesting?

Think about the last time something genuinely grabbed your attention. Not because you had to pay attention to it, but because you couldn't help it. Maybe it was a throwaway comment that opened up a question you'd never thought to ask. A detail in a documentary that suddenly connected two things you'd been holding separately for years. A problem at work that you kept turning over in your head long after you should have stopped. That's interestingness. And it's a surprisingly hard thing to pin down because of how personal it is.

The same puzzle that one person finds fascinating will strike someone else as pointless. Interestingness isn't really a property of the thing itself, it's something that happens in the gap between the thing and the person encountering it. It depends on what you already know, what you're currently mulling over and what gaps you're carrying around without quite realising it. You're not consciously doing all this.

There's also something almost physical about it. Genuine interest has a pull to it. Time moves differently. You lean in rather than away, it becomes motivational and shapes what you do next. And you don't always know in advance what will interest you, sometimes the thing that grabs you most is the one you nearly walked past.

All of which makes for a fascinating problem for AI research. If you're trying to build a system that can keep learning on its own you need it to have some sense of what's worth pursuing and what isn't. You need it, in some way, to find things interesting.

Jeff Clune and his collaborators have been working on exactly this. Their framework, OMNI (Open-endedness via Models of human Notions of Interestingness), puts the problem clearly: even after you filter for tasks an AI system can actually learn, you're left with a vast number that are technically learnable but essentially pointless, maybe small variations of things already done or dead ends. The system has no way of telling the difference between a task worth pursuing and one that isn't.

Their solution is to use a large language model as a guide leaning on the LLM's internalized sense of what humans find interesting to steer the agent toward more worthwhile territory. A follow-up framework, OMNI-EPIC, takes this further still, having foundation models generate entirely new learning environments in code rather than just selecting from existing ones. The research is active, serious, and genuinely creative.

But interestingness doesn't always wait to be engineered. A new paper by Clune and collaborators, AI Finds a Way, documents something stranger: AI systems that discover unexpected, unintended solutions entirely on their own. Not because they were guided toward interesting territory, but because reward-driven optimisation led them somewhere nobody anticipated, exploiting loopholes, circumventing design constraints, and occasionally stumbling onto previously unknown scientific phenomena. It's a kind of curiosity that emerges uninvited, from the bottom up rather than the top down.

It's a striking contrast. One paper tries to teach AI what's worth finding. The other documents AI finding things nobody asked for. And yet in both cases the question of what makes something genuinely interesting to a person remains unanswered. Which is exactly what I keep coming back to.

When you use a language model as a stand-in for human interestingness, whose interestingness are you actually capturing?

A language model learns from what humans chose to write down, share, and come back to. That's an enormous amount of text but it isn't a neutral sample. It skews toward things that were expressible, shareable, and already validated by enough people to make it onto the internet. The private obsessions, the niche fixations, the things you found fascinating at 2am that you'd struggle to explain to anyone do not register. What the model absorbs is closer to a statistical average of expressed human interest, flattened across millions of people and contexts into a single signal.

The corpus also carries familiar cognitive biases, now operating at scale. Survivorship bias means the model learns almost nothing from ideas that were explored and abandoned and only from what was interesting enough to survive into text. Bandwagon effects mean topics that gain traction get written about more, crowding out quieter signals. Status quo bias tilts the whole thing toward established thinking rather than the questions that challenge it. None of this makes the approach wrong. But the limitations are worth understanding, not dismissing.

There's also something about human interestingness that's genuinely hard to capture from the outside. That pull, that feeling of time disappearing, the surprise of finding yourself absorbed in something you nearly walked past it's relational, personal, and rooted in the specific person you are right now. A proxy can recognise the pattern. It can't replicate the person.

Clune knows this is hard. He's called interestingness the "Achilles Heel" of open-endedness research, and OMNI is his best attempt at answering it.

What I keep coming back to isn't really about the technology. It's about what we lose, or don't, when something as personal as curiosity gets averaged out. I don't have a clean answer to that. But I think it's a question that deserves to be asked more often.

Thought pondered by Sarah exploring the intersection of AI, creativity, and human wellbeing